Software, commissioned.

Amirus designs, builds and runs AI‑powered platforms and the cloud infrastructure beneath them, from first architecture to production at scale. Nothing goes live until one of our engineers signs the handover.

  • AI & ML
  • Cloud & deployment
  • Web & mobile
  • Product design

Since 2020: 70+ projects shipped for 50+ clients, and 86% of them came back for the next one.

AM‑2611 · commissioning record Simulated
System schematic Web, mobile and point-of-sale clients route through an edge layer and API gateway to three services, which use an LLM, a vector database and Postgres. Every node passes its checks. Web Mobile POS Edge API Agent Orders Stock LLM Vectors Postgres
Handed over by Lead engineer · 14:02

Architecture to on-call, end to end.

  1. AI & ML integration

    LLM features, retrieval over your own documents, agents that call your systems, forecasting and computer vision. Wired into the product you already have, with evaluations, guardrails and cost limits.

    You receive eval suite, prompt & model registry, cost dashboard

  2. Cloud & deployment

    Kubernetes, CI/CD, multi-region and multi-tenant SaaS, on‑prem and air‑gapped installs, large migrations off legacy stacks. Observability first, so we know before your customers do.

    You receive infrastructure as code, runbooks, alerting, rollback plan

  3. Web, mobile & SaaS

    Customer platforms, enterprise systems, CRM, ERP and POS, e‑commerce and logistics software, built to carry real load. React, Next.js, Flutter, React Native, .NET, Python.

    You receive source code, test suite, store-ready builds, docs

  4. Product & interface design

    Research, UX, design systems and AI‑native interfaces. Designed around how people actually work, then checked for accessibility before anyone writes code.

    You receive prototypes, design system, accessibility audit

AI that shows its work.

A model that sounds right is easy. One that cites its sources, admits uncertainty and stays inside a budget takes engineering. We ship every AI feature with these instruments, the same ones behind the order-taking assistant we built for Salesleon.

Recorded trace
Pick a question
Retrieved from your documentsscore

    Answer

    Releases without a war room.

    We run your software the way we run our own: watched, scaled and rolled back by the pipeline, so a bad release costs seconds instead of a weekend.

    A long aisle between black server cabinets in a data centre, lit from above
    Server aisle in a data centre. We deploy to public cloud, colocated racks and on‑prem clusters alike.Photo: Christopher Bowns, CC BY-SA 2.0

    Watch a release go out.

    Scroll to step through one Tuesday in production.Tap a step, or let it play.

    1. Three replicas of v1.14 serve normal Tuesday traffic. Every service reports latency, errors and saturation to the same dashboard your team sees.

    2. A campaign lands and requests triple. The autoscaler adds pods before latency moves. Nobody gets paged.

    3. v1.15 takes 10% of traffic. Its error rate crosses the SLO, so the pipeline rolls it back on its own in 38 seconds. On‑call gets the trace, customers get nothing.

    4. The patch, v1.15.1, passes its canary and goes to 100% in stages. The change log, the trace and the sign-off sit together on one page.

    prod · eu‑west Simulated
    Ingress
    Requests/s1,180
    p95 latency84 ms
    Error rate0.02%
    Canary—
    14:00:02 v1.14 healthy · 3/3 ready
    A rack of ten identical white server nodes stacked floor to ceiling
    A ten-node cluster in a rack: what on-prem usually means.

    Your cloud, or no cloud at all.

    The same release pipeline ships to a multi-tenant SaaS cluster or to a rack in a warehouse with no internet. Switch it and see what changes.

    • Images from a managed registry, pulled per tenant
    • Secrets in cloud KMS, rotated every 30 days
    • LLM calls through a gateway with per-tenant budgets
    • Updates continuous, canaried per region

    Recorded pipeline log

    ok build      api@a41f9c2          1m12s
    ok test       1,284 passed         2m40s
    ok scan       0 critical, 2 low    0m38s
    ok image      sha256:9e1b..c07d    0m21s
    ok staging    smoke 42/42          1m05s
    .. canary     10%, watching SLO    0m00s
    -- promote    25 > 50 > 100%
    -- sign-off   lead engineer

    Salesleon, built for a client.

    We built Salesleon for a client: an AI sales inbox, a POS and a warehouse system for e-commerce sellers in Bangladesh. One WhatsApp order, three systems: follow it from the chat to the shelf.

    Salesleonsalesleon.ai

    WhatsApp · new chat

    2 basmati 5kg and 1 olive oil, to Uttara please

    Done. Tk 3,450 cash on delivery, arriving tomorrow.

    Fraud check: low riskCourier assigned

    The AI takes the order on WhatsApp, Messenger, Instagram or TikTok. Prices, stock and fraud rules stay in plain code.

    Salesleon POSpos.salesleon.ai

    Store 03 · Order A‑2611
    via Salesleon · Tue 14:06

    • Basmati rice 5kg ×22,600
    • Olive oil 1L850

    Total (Tk)3,450

    Cash on delivery · packed

    The shop rings it up and packs it.

    Salesleon WMSwms.salesleon.ai

    Bin C‑14 · rice 5kg · 41 left

    Stock leaves the shelf, and the forecast flags the reorder before it runs out.

    Our games studio, and our sister company.

    • Amirus Studio

      Android games, built and live-operated in-house.

      Our games studio. Neon Dash: Boss Runner 3D is live on Google Play, with Jarvolution and Sharkco: Ocean at War in testing.

    • Reach Goal

      Our sister company in Dhaka.

      A full-service marketing agency: social media, Meta, Google and TikTok ads, branding and websites for Bangladeshi brands. When a product needs customers, they run the launch.

    How a project gets commissioned.

    Five phases, each with something you can hold at the end and a person who signs for it.

    DonePhaseWhat you receiveSigned byTypical length
    Discover Problem statement, constraints, a fixed-price estimate for the next phase Your product owner 1–2 weeks
    Architect System design, data and AI plan, clickable prototype, delivery schedule Our lead engineer and you 1–3 weeks
    Build Working software every two weeks on a staging URL you can use Sprint review, both sides 8 weeks to multi-year
    Commission Load test, security scan, AI evals, runbooks, go-live and rollback plan Our lead engineer 1–2 weeks
    Operate Monitoring, on-call, monthly reports, or a clean handover to your team Whoever holds the pager Ongoing

    Start a project.

    Tell us what you're building, from a single AI feature to a platform programme. An engineer, not a salesperson, replies within one business day.

    1. Within one business day

      An engineer reads your work order and replies with questions, not a brochure.

    2. A 30-minute call

      We talk through the problem, what exists today and where it has to run.

    3. A fixed-price Discover phase

      One to two weeks, a written estimate for the build, and no lock-in.

    Work orderAM‑WO‑····
    What are you building?
    Where must it run?

    Tell us who to reply to.

    We need a valid email to reply.

    We reply from an inbox, not a CRM. No newsletter, no sharing.